A Dynamic Target Visual Positioning Method Based on ROI

The method of visual positioning can be mainly divided into fixed camera system and mobile camera system. In this paper, we propose a dynamic target positioning method based on ROI (regions of interest), which utilizes the deep learning method to detect targets and employs the fixed camera system to locate the targets. The ROI method proposed only process the region of target, which can reduce the time-consuming, and it can solve the problem that none or less feature points of the target is detected in 3D reconstruction. We make a dataset of the experimental car and use YOLOv2 to train the dataset, by which the training model of the experimental car is obtained; then the trained model is used to detect the experimental car in the video data which acquired by two USB cameras and get the ROI of the moving target. According to the triangulation method, only the ROI of the image data at the same time is reconstructed, and the average of the obtained coordinates as the position of the car at that moment. In the experiment, we use the positions obtained by optitrack system as the true values, and compare the positions got by the method of this paper (ROI method) with the true value. The experimental results show that the ROI method proposed can be used to locate the dynamic target with the positioning accuracy at the cm level.